#Images generating from an init image, despite being turned off
1 messages · Page 1 of 1 (latest)
uh... what?
gonna need a lot more info to know what's going on
can you at least screenshot the whole UI, showing a completed gen and all the params and etc
so, (A), the generation you have displayed there is not your current one, prompt doesn't match
(B) do you possibly have a static seed set and that's what's throwing you off?
(C) you have a giant prompt, model with loras, etc. ... are you just getting similar images because you're generating about the same thing over and over? The more you specify in prompt and all, the less variation the output will have, naturally
(D) you really love the , character don't you lol
I tried with different prompts to try out then switched back to the original one, even with different background and details
the seeds is randomized
is that a bad thing? hahaha
not a big deal just noticeable
eg this , , , in the middle is wonky
and the , at the start of a segment doesn't quite make sense
nor the comma at the end before the segment
I usually put loras in between
loras also don't need commas between them lol
loras aren't part of the prompt, there's just prompt syntax to put them in for compatibility with auto-style usage
oh also do note, your segment prompts are not great
eg ((character design)) as an entire prompt is weird
especially given your usual prompt is a mile long
you'd probably want at least the character name included there i'd imagine
and facial details like hair/eye color/etc
... if you can figure out what color the eyes are meant to be lol
oh no i see in the image history filename it's a different seed so not that
character is actively moving farther/closer between gens
so probably ^ C, just... between the model and loras and the giant prompt, you've specified so much that it doesn't see much reason for variation
model might also be a bit burnt on that character potentially
hm, I see, I usually do those long prompts because the Lora ain't great since I have very few images of each char
oh it's your own lora?
yeah quite possibly you burnt it then
ie you trained it for too long / too high of LR, on too few images, so it's learned to just spit out the images you gave it rather than trying to learn who the character is
The point is that before I used an init image there was enough variety
there's definitely not an init image there
and there's no magic ai memory, certainly not anything that could survive a reinstall either lol
you might go back and find the images that had better variety, and see if you had different params or prompt setup that worked better on those